AssemblyAI Voice Agent API at $4.50 an hour all-in

AssemblyAI's Voice Agent API is listed at $4.50 an hour, which is $0.075 a minute, and the price includes speech-to-text, the LLM, text-to-speech and orchestration. A 10,000-minute month at that rate is $750. One detail changes your budget: streaming is billed on session duration, not on how much audio was spoken.

What the page lists

From AssemblyAI's pricing page, read 2026-10-05, with Cartesia's agent price from its own page for context.

Session time, not audio time

The page says streaming is billed on session duration. A session that sits open while the caller thinks, or while the line is silent, costs the same per minute as one full of speech. For a support line with long pauses, the effective price per minute of actual speech is higher than $0.075. Model it with your own call logs: billable minutes equal call length, not talk time.

Hold times, voicemail greetings and ringing that your system keeps open all count if the session is open. Closing sessions promptly is the cheapest optimisation.

What a month costs

Arithmetic on $0.075 a minute. It assumes every minute of call time is session time.

All-in versus assembling the parts

A bundled price is simple but hides the split. Building your own agent means paying separately for streaming STT, an LLM, TTS and orchestration, as with OpenAI's gpt-realtime-2.1, whose page lists per-token audio prices ($32 input and $64 output per million audio tokens). The bundled number is easy to forecast; the unbundled one is easier to tune. The sources for this post give no latency or quality figures for the Voice Agent API, so test a call before you commit.

Sume does not offer a voice-agent product. Its audio services are asynchronous jobs: Cartesia Sonic text-to-speech, speech-to-text on video clips, music and lip-sync. They suit produced content, such as an ad read or a talking-head clip, not live calls.